Predicting fillet lipid composition of Oncorhynchus mykiss using different statistical tools

Lipid composition of animal products is pivotal for human health, seafoods are not an exception. Objective of this study was to benchmark different sources of information and statistical tools to predict fillet lipid composition in rainbow trout (Oncorhynchus mykiss). The fatty acids (FA) profile of aquafeeds and fish flesh were determined by gas-chromatography. A total of 228 records on single-fish analyses was available. Fish were fed 25 different diets and housed in 77 tanks. The dataset was assembled from five trials. Fish diets were characterized for their chemical composition, ingredient proportion as well as FA composition, mainly docosahexaenoic (DHA, C22:6n-3), eicosapentaenoic (EPA, C20:5n-3) acids, the total of polyunsaturated FA omega-3 (n-3 PUFA) and the sum of EPA+DHA. These sets of variables were used as predictors. Statistical models included multiple linear regression with shrinkage (MLR), partial least square regression (PLS) and random forest (RF). A three-fold cross-validation was set up, training on 2 fish samples from each tank and validating on the remaining samples. Pearson correlation between predicted and measured values in the validation set was used as a measure of prediction accuracy, averaged over the three replicates. Results show predictions based on the proportion of ingredients and the diet’s chemical composition achieved accuracy of at least 0.75 for the studied fillet characteristics. As an alternative, the diet’s lipid composition can be used alone. The use of all sets of predictors in the same model seemed to be redundant. Random forest outperformed the other statistical tools, suggesting that the predictors act in interaction rather than marginally. Results are encouraging and the quality of trout fillet could be predicted leveraging on diet description, with some degree of accuracy. Further studies will focus on assessing the predictive ability of the same models when a specific diet composition is not being studied and yet included in the training set.